Integrated Regime-Aware Wind Power Forecasting using Multi-Altitude Meteorological Features and Hybrid Machine Learning

Abstract

Accurate wind power forecasting is critical for grid stability and renewable energy integration, yet the inherent variability of atmospheric conditions presents significant challenges. This study proposes a unified and scalable pipeline that integrates wind regime detection, temporal sequence modeling, and regime-conditioned deterministic and probabilistic power forecasting. Using 8 years of high-resolution meteorological data from multiple altitudes, we engineer a comprehensive set of physically interpretable features, including wind shear, temperature gradients, and rolling statistics. Regimes are identified via KMeans and Gaussian Mixture Models, with Principal Component Analysis applied post-clustering for visualization and interpretation. Temporal regime dynamics are characterized through both empirical and Markov transition matrices and modeled using Long Short-Term Memory (LSTM) networks for regime sequence prediction. For power forecasting, regime-specific models are developed using tuned ensemble regressors (XGBoost, LightGBM, CatBoost, and Random Forest), complemented by probabilistic approaches including Quantile Regression Forests, quantile-based XGBoost, and Bayesian neural networks. Results show that regime conditioning significantly enhances forecasting performance, with the stacked meta-learning ensemble achieving R2 = 0.997 and over 30% reduction in MAE compared to baseline methods. Probabilistic models produce well-calibrated prediction intervals, providing uncertainty-aware forecasts suitable for operational decision-making. This work contributes a novel end-to-end framework that jointly models regime persistence, transitions, and regime-conditioned power output, incorporating uncertainty quantification through Quantile Regression Forests, quantile-based XGBoost, and Bayesian Neural Networks, thereby bridging a gap in the literature where these components are often treated in isolation. The approach advances both accuracy and interpretability, offering practical value for wind farm operation and renewable energy integration.